I like this better for three reasons:
- Since usually it is scripts producing this JSON - we have the most control at that point for determining the failure and we don't have to deal with an artificial dependency between the tool's stdio and the output.
- At some point we could potentially allow some datasets to be ok now even though the job fails.
- It is a cleaner interface at the Python level between job finish and output collection IMO (no need for isinstance checking).
John Chilton
121285b40b